People of the Levante

About the Project

The eastern Mediterranean of the 11th–14th centuries was a region of extraordinary cultural, linguistic, and religious diversity. Yet traditional prosopography of the Crusades era has overwhelmingly focused on ecclesiastical dignitaries, ruling families, and military orders — the groups with the best documentary survival.

This project asks: can generative AI and Knowledge Graphs make the full population of the medieval Levant visible at scale? That means not only knights and bishops, but refugees, artisans, merchants, interpreters, women, children, and unnamed collectives who appear only in fragments of surviving sources.

The pipeline combines two layers of intelligence. A large language model (Google Gemini) extracts person-like signals from historical texts — names, titles, epithets, roles, group markers — while a knowledge graph of 126 curated authority records grounds those signals in established scholarship. Mismatches between the two layers are treated as data, not errors: they reveal ambiguity, missing entities, and the limits of current authority files. Historical scholars adjudicate the results through a Human-in-the-Loop review interface.

⚗️ Proof of concept. This is a working demonstration of the methodology, not a finished research database. The authority file currently contains 126 persons; the source corpus is small. Both will grow iteratively through scholar contributions.

How It Works

Historical texts (PDFs or plain text) are ingested from the source corpus.
Gemini extracts person-like mentions with context: name, title, toponym, role, collective vs. individual, confidence score.
Each mention is fuzzy-matched against the authority file of known crusader-era persons, producing ranked candidates with confidence scores.
Scholars review matches in the Explorer — accepting, rejecting, or flagging candidates with optional comments.
Exported decisions feed back into the next iteration, refining the model.

Layer 1 — LLM Extraction

Google Gemini reads historical texts in multiple languages and identifies person-like signals, including unnamed collectives and ambiguous references. Uncertainty is kept explicit.

Layer 2 — Knowledge Graph

A curated authority file of 126 crusader-era persons (sourced from an Omeka database) provides the scholarly substrate. Fuzzy matching surfaces ranked candidates with scored confidence.

Human-in-the-Loop

Scholars accept, reject, or flag each candidate link. Ambiguity is a feature, not a bug: contested links expose gaps in the authority file and guide future curation.

Open & Citable

All source code is public. Extracted data and adjudication decisions are downloadable as JSON. BibTeX references are generated per document.

Project Team

Jochen Burgtorf — Professor of History, California State University Fullerton. Expert in medieval Mediterranean history and digital prosopography.

Tobias Hodel — Associate Professor of Digital Humanities, University of Bern. Specialist in AI methods for historical documents, HTR, and data modeling.

Laura Morreale — Independent Scholar, Harvard University. Research focuses on medieval cultural contact, multilingual relations, and the Eastern Mediterranean.

A book chapter co-authored by Burgtorf and Hodel (ed. Morreale) outlines the theoretical framework. This repository is the technical proof of concept for a grant application.

Contribute a Source

Do you work with primary sources from the medieval Levant? You can submit a PDF or plain-text document for processing. Uploaded documents are reviewed before entering the pipeline — you will be notified of the outcome.

By submitting a file you confirm that you hold the necessary rights to share it for research purposes. Only PDF and plain-text files up to 50 MB are accepted.

Submit a document →

Source Code & Data

Everything is open source. Pipeline code, authority file, and extracted data are available at github.com/thodel/outremer.